Identification of DNA N6-methyladenine modifications in the rice genome with a fine-tuned large language model

DNA N6-methyladenine (6mA) plays a significant role in various biological processes. In the rice genome, 6mA is involved in important processes such as growth and development, influencing gene expression. Therefore, identifying the 6mA locus in rice is crucial for understanding its complex gene expression regulatory system. Although several useful prediction models have been proposed, there is still room for improvement. To address this, we propose an architecture named iRice6mA-LMXGB that integrates a fine-tuned large language model to identify the 6mA locus in rice. Specifically, our method consists of two main components: (1) a BERT model for feature extraction and (2) an XGBoost module for 6mA classification. We utilize a pre-trained DNABERT-2 model to initialize the parameters of the BERT component. Through transfer learning, we fine-tune the model on the rice 6mA recognition task, converting raw DNA sequences into high-dimensional feature vectors. These features are then processed by an XGBoost algorithm to generate predictions. To further validate the effectiveness of our fine-tuning strategy, we employ UMAP(Uniform Manifold Approximation and Projection) visualization. Our approach achieves a validation accuracy of 0.9903 in a five-fold cross-validation setting and produces a receiver operating characteristic (ROC) curve with an area under the curve (AUC) of 0.9994. Compared to existing predictors trained on the same dataset, our method demonstrates superior performance. This study provides a powerful tool for advancing research in rice 6mA epigenetics.